Agent contracts reference - Harmonia

Python contracts live in agent/harmonia_agent/agent_models.py; shared application representations live in src/lib/types.ts. Pydantic models forbid unknown fields and deterministic validators reject ungrounded or authority-expanding output.

Agent	Input	Output	Boundary
Nimi	`AnalystInput`	`SourceAnalysis`	Evidence analysis only
Ryan	`StrategistInput`	`StrategistResult` containing `ContentStrategy`	Strategy and source-grounded briefs only
Temi	`EditorialPlannerInput`	`EditorialPlan`	Operational editorial plan; supported channels only
Noni	`CopywriterInput`	`ContentDraft`	Final-copy drafting, not approval or publishing
Dara	`EditorialReviewInput`	`EditorialAssessment`	Bounded acceptance or revision issues
Maya	Typed surface context	`SurfacePlan`	Presentation plan, no durable mutation
Nova	`LiaisonInput` plus read tools	`LiaisonAnswer`	Read-only explanation and navigation

Evidence identifiers form a closed world per invocation: references not supplied in the input are invalid. Memory facts retain Firestore provenance and remain advisory. See the agent team for role-by-role detail.

Durable handoffs

From	To	Exact handoff	Deterministic gate
Nimi	Ryan	Persisted `SourceAnalysis` plus canonical digest	Evidence and provenance validation
Ryan	Temi	Persisted `ContentStrategy`, lineage, version, and digest	Authenticated strategy approval bound to the current digest
Temi	Noni	One selected `EditorialPlan` item, exact Ryan brief, referenced Nimi evidence	Plan validation, persistence, and deterministic item selection
Noni	Dara	Immutable `EditorialReviewInput` and exact `ContentDraft`	Draft contract and claim-ledger validation
Dara	Noni or final approval	Structured issues or the exact accepted draft	One-revision limit and assessment validation
Maya	UI host	Reference-only `SurfacePlan`	Entity and component validation, then server hydration
Nova	Operator UI	Read-only `LiaisonAnswer`	Skill/tool trace and evidence-citation validation

Evidence namespaces

Basis	Valid identifier source	What it may support
Operator context	IDs in typed company, product, audience, objective, and campaign input	Supplied business assertions
Source analysis	Nimi moment and angle IDs	Source-derived claims and Ryan briefs
Performance	Verified performance observation IDs	Performance-derived recommendations
Memory Bank	Eligible fact IDs with Firestore provenance	Advisory cross-job learning only
Public research	Request-bound search evidence IDs with validated native grounding metadata	Current external claims within the exact research request
Internal reads	Tenant-scoped tool evidence IDs or planning snapshot identity/digest	Prior publication, job, engagement, capacity, dependency, and calendar-projection facts

All models forbid extra fields. Unknown references, empty grounding, skill references used as facts, memory used as authorization, and role-expanding claims fail closed. See Agent skills, tools, prompts, and traces for the runtime inventory.

Last modified on August 28, 2026